When I code with Claude Code, I use it through Orca. If you look at the Orca settings, you can see the total tokens the agent has used. To give you an idea of how much AI I typically use, I consumed 0.7 billion tokens in just one week. That means I used 700 million tokens. This number includes cache tokens, so if I calculate the actual meaningful tokens I used, it would be a bit less. But anyway, I am personally burning through around 700 million tokens a week. There are obviously people who use way more than me, but I still think I am someone who uses agents pretty heavily. Coming from that perspective, I often feel disillusioned when I study AI lately. The story I want to talk about today is whether there is really any need to study how to handle AI well. It might sound obvious, but I constantly think AI is moving too fast, like a passing fad.
 |
| The reality of burning through 700 million tokens in a single week |
The trends and themes change entirely too fast. When GPT first came out, we only had to use it on the web, but then it popped up as an API and a CLI. Then came prompt engineering, context engineering, harness engineering, OpenClaw automation, orchestration, loop engineering, and graph engineering. I just named what comes to mind right now, and there is already a huge pile of them. The bigger problem is that the speed at which new things appear is only getting faster. I feel like the fatigue is just stacking up. When OpenClaw first dropped, everyone was going crazy trying to set it up. But now mobile integration is just a basic default. Connecting Gmail or documents takes exactly one click. Now I wonder why we even need OpenClaw at all. Things like automation are attached to Codex by default.
I have been using Orca for about a month now while it is trending. If you install Orca, you immediately get harness, skills, various engineering methods, computer use, and mobile integration right out of the box. The things we scrambled to study and configure to use AI better just become basic default services a few months later. Things getting easier is obviously a good thing. But watching this flow continue for almost two years brings an overwhelming sense of emptiness more than anything else. Whenever a new technology comes out, I just keep thinking about why I should even bother studying it right now.
Raw Performance Overwhelms Everything
I mentioned this in a piece I wrote a few days ago, but my old coworker who left the company three years ago told me he is developing a game with an acquaintance. The game is obviously still in development. I talked with my friend recently, and he told me he actually did almost nothing while making this game. He just dropped in the 3D assets at the very beginning and that was it. After that, the UI work, illustrations, user ranking board, and game logic were all finished entirely using vibe coding. He has never even seen how the code was implemented and says he has never opened a code editor once. He did not apply any harness or anything to the AI he assigned the work to. He did not use a single engineering skill and just did raw vibe coding, but the results came out in a single day.
If you want to read more about how this same friend previously struggled by overcomplicating his process and obsessing over tools you should check out my recent piece on My Developer Friend Wasted Three Years Chasing Perfection.
What I realized is that this would have been flat out impossible just last year. If you wanted to do something like this back then, you had to do intense harness engineering, building code templates, defining how to build each feature, standardizing SOLID principles, and documenting exactly what algorithms to use in md files. You had to spoon feed every single detail yourself just to get a usable vibe coding result. So, up until last year, this kind of engineering work was genuinely required. But now I really wonder if any of that is still necessary.
 |
| Thinking you are holding the harness when AI is actually pulling you |
The conclusion is that the improving performance of the AI itself is simply everything. Raw performance improvement simply overwhelms everything else. Applying things like harnesses to an already improved AI will naturally yield better results. But I honestly wonder what the point of that even is. New techniques like graph engineering are trending right now, but I question what meaning they truly hold. I have developed a bit of skepticism about what it means to study AI handling techniques. Ultimately, much better agents will emerge in a few months based purely on raw performance, so studying these complex techniques right now just for a tiny performance bump feels pointless. When these thoughts expand, I start wondering where AI is ultimately going to end up. I do not want to preach about the future of AI. But the trends change so fast, and I question their utility, so my desire to learn keeps dropping.
The Irreplaceable Value of Real Experience
Despite all of this, I still believe there is something that really matters here. Computer science knowledge and actual hands on development experience still seem crucial. As vibe coding becomes more established, people who have actually built things purely on their own will become even more rare and valuable. I think this is a very important point. Let me give you a simple example, because a similar situation happened when my friend was making his game.
 |
| Why basic Unity knowledge easily beats complex AI calculations |
His game requires fish to constantly swim back and forth in front of the camera. He asked the agent to build that feature, and it started implementing it by calculating and inferring the fish's allowed movement range based on the camera position and field of view. If you are someone who has never used Unity directly, or has only ever done vibe coding, you might actually think this is the correct way to implement it. But anyone who has actually used Unity knows how incredibly inefficient this is. They would immediately wonder why you would ever do such useless calculations. They know the simplest method is to just create a collider, place it right in front of the camera, and make the fish swim only inside of that box. If you do it that way, you eliminate the need for calculations, and level design becomes so much easier during development.
But this kind of method has to be conceived by the person handling the agent, the pilot. You have to look at the method the AI suggests, tell it that it is a bad approach, and direct it to do it the other way. Only then will the AI agree that your way is better and correct its development direction. If you lack the knowledge and purely rely on vibe coding, the AI is not going to automatically land on the most appropriate method on its own. In short, I believe this is why two people can use the same AI and end up with wildly different results, one great, one terrible.
This perfectly illustrates why simply writing syntax is no longer the main job of a programmer. I wrote a deep dive on this shift in AI Proves Real Developer Skill Was Never About Code where I explain why architectural thinking and problem solving are the only things that truly matter now.
I truly believe the broader the experience and vision of the person handling the agent, the better the direction the AI will take. But this kind of capability is a skill you can only acquire if you have the experience of developing directly without any AI or agents. This specific gap is going to create a massive wall in the end. My friend lacked that development experience, so he started off very inefficiently. But he told me he eventually revised it using the methods of his acquaintance, who actually had game development knowledge, and saw much better results. So ultimately, I think this kind of raw experience is just too important.
Of course, if much more time passes and AI develops even further, I think it will eventually resolve even these kinds of bottlenecks. For now, my honest bet is that the person who has actually built something with their own hands still has the edge, and I plan to keep betting my own time on that. But if the day comes when AI effortlessly bulldozes through even that final wall of human intuition, we have to ask ourselves if we just found the ultimate solution, or if we simply became the problem.
For deeper reads on tech, business, and the shifting economy, follow The Techtonic for weekly breakdowns.
Latest Insight: [Why Every Indie Game Success Feels The Same Lately]
Trending Insight: [Stop Learning and Just Start Copying in the AI Age]
Comment